Machine Recognition in Complex Domain - High Dimensional Neurocomputing: Growth, Appraisal and Applications - page 153

Figure 7.6 shows that the feature extraction criterion performs reasonably good, when
48 features of subspace are chosen. The increase in subspace dimension does not
improve the overall performance. Though recognition rate in R PCA increases slowly
beyond 48 features, but still it is less than C ICA.

7.5.2 Performance in Indian Face Database

This experiment considers 500 color images corresponding to 50 subjects of Indian
face database [ 63 ]. Each image is of size 640 480 pixels. All images have a bright

homogeneous background and different poses and emotions. For each individual,
following poses have been included for the face: looking front, looking left, looking
right, looking up, looking up toward left, looking up toward right, and looking down.
In addition to the variation in pose, images with four emotions such as neutral, smile,
laughter, and sad/disgust are also included for every individual. For example, few
sample images are shown in Fig. 7.7 . The number of images considered for training
and testing is 200 and 300 respectively, comprising of all subjects.
Eigenvectors with higher eigenvalues provide more information on the face varia-
tion than those with smaller eigenvalues. The eigenvectors (eigenfaces) correspond-
ing to largest eigenvalues, computed from the Indian face database are shown in
Fig. 7.8 . The eigenvectors are ordered according to decreasing eigenvalues. Each
individual face in the face set can then be approximated by a linear combination
of the eigenvectors associated with largest eigenvalues. Based on the argument that
for tasks such as face recognition much of the important information is contained
in high-order statistics; it has been proposed [ 31 , 38 ] to use ICA to extract features
for face recognition. The basis vectors (basis images) obtained by the Infomax algo-
rithm for ICA representation are shown in Fig. 7.9 . We obviously want to capture
as much variations as possible of the training set with as fewer number of subspace
dimensions. The graph in Fig. 7.10 allows us to see more clearly how much variation
is captured by number of eigenvectors. This information prompted us to select 60
eigenvectors from the set of training images of 50 subjects for further feature extrac-
tion. They correspond for keeping more than 91%of total variance in the eigenvalues
of this dataset.
The test results of Indian face database with different feature extraction techniques
and different neural classifiers are presented in Table 7.2 in terms of different mea-
sures. It again reveals the superiority of C RSP-based classifier. The feature vectors

×

Fig. 7.7 Some example images from the Indian face database